Discovery and validation of candidate host DNA methylation markers for detection of cervical precancer and cancer

Abstract

Human papillomavirus (HPV) testing has been recently introduced as an alternative to cytology for cervical cancer screening. However, since most HPV infections clear without causing clinically relevant lesions, additional triage tests are required to identify women who are at high risk of developing cancer. We performed DNA methylation profiling on formalin-fixed, paraffin-embedded tissue specimens from women with benign HPV16 infection and histologically confirmed cervical intraepithelial neoplasia grade 3 (CIN3), and cancer using a bead-based microarray covering 1,500 CpG sites in over 800 genes. Methylation levels in individual CpG sites were compared using a t-test, and results were summarized by computing p-values. A total of 12 candidate genes (ADCYAP1, ASCL1, ATP10, CADM1, DCC, DBC1, HS3ST2, MOS, MYOD1, SOX1, SOX17, and TMEFF2) identified by DNA methylation profiling, plus an additional three genes identified from the literature (EPB41L3, MAL, miR-124) were chosen for validation in an independent set of 167 liquid-based cytology specimens using pyrosequencing and targeted, next-generation bisulfite sequencing. Of the 15 candidate gene markers, 10 had an area under the curve (AUC) of ≥ 0.75 for discrimination of high grade squamous intraepithelial lesions or worse (HSIL+) from <HSIL cytology using at least one assay. Overall, SOX1, DCC, and EPB41L3 showed the best discrimination with AUC values of ≥0.80, irrespective of methylation detection assay. In addition to verifying candidate markers from the literature (e.g., SOX1 and EPB41L3), we identified novel markers that may be considered for detection of cervical precancer and cancer and warrant further validation in prospective studies.

Introduction

High-risk human papillomavirus (HPV) testing is emerging as a preferred alternative to current cytology-based methods for cervical cancer screening, as it demonstrates improved sensitivity for detection of cervical cancer precursor lesions (cervical intraepithelial neoplasia grades 2 and 3 (CIN2 and CIN3)) compared with cytology alone. However, since a majority of HPV infections are transient and spontaneously clear without causing clinically relevant lesions, the specificity and positive predictive value of HPV DNA testing for cervical precancer and cancer is relatively low¹. Thus, additional triage tests are needed in order to distinguish women with prevalent high-grade disease who are in need of immediate referral to colposcopy (and ultimately treatment), from those with benign cervical HPV infections².

Simple molecular-based approaches, such as DNA methylation testing of host genes in cytological specimens, offer promise as an objective biomarker for cervical cancer screening and triage³. Such assays are robust, have the potential for various degrees of automation, and can be readily applied to a variety of specimen types, including formalin-fixed, paraffin-embedded (FFPE) tissues and self-collected cytology samples². Previous studies have identified increased DNA methylation of host tumor suppressor genes associated with cervical precancer and cancer³. Some of the most promising markers, including CADM1, MAL, miR-124, SOX1, TERT, and EPB41l3, have been evaluated extensively in cervical cancer screening studies, and have demonstrated equivalent performance as cytology for triage of HPV-positive women for the detection of precancers and cancers^4–13. However, most studies conducted to-date have been based on a candidate gene approach^{5, 7, 8, 14–16}; and very few unbiased DNA methylation profiling studies have been conducted, particularly those that aim to distinguish precancer (i.e., CIN3) from normal tissue^{4, 14, 17–22}. Likewise, efforts to compare the results of different assays have been limited, with a majority of studies utilizing quantitative methylation-specific PCR (qMSP) for methylation detection³.

Thus, the present study aimed to identify candidate host DNA methylation markers for detection of cervical precancer and cancer, using an objective, bead-based microarray analysis of FFPE tissue specimens. We then sought to validate the top candidate markers in an independent set of liquid-based cytology (LBC) samples using two high-resolution DNA methylation detection assays: pyrosequencing and targeted, next-generation bisulfite sequencing (tNGBS), to assess their potential as triage biomarkers for cervical precancer and cancer.

Materials and Methods

Study Population for Biomarker Discovery Using a Bead-Based DNA Methylation Array

FFPE tissue specimens were obtained from women enrolled in the Study to Understand Cervical Cancer Early Endpoints and Determinants (SUCCEED), a large study of women 18 years of age or older with an abnormal Pap smear who were referred to colposcopy or treatment at the University of Oklahoma Health Sciences Center (OUHSC) between 2003 and 2011^{23, 24}. Written informed consent was obtained from all women enrolled in the study and Institutional Review Board approval was provided by OUHSC and the U.S. National Cancer Institute.

Participants provided specimens for ThinPrep® Pap testing (Hologic, Inc., MA) and LINEAR ARRAY® HPV Genotyping Test (Roche Molecular Diagnostics, CA). Colposcopy and specimen collection procedures for SUCCEED have been described elsewhere^23–25. Briefly, all histologically-confirmed CIN3 and a majority of CIN2 lesions were treated by loop electrosurgical excision procedure (LEEP) of the transformation zone, as per standard practice. Every LEEP specimen was manually dissected into 12 topographically designated segments for detailed histopathological mapping. Of the 12 segments, 10 were subsequently FFPE and histology results were determined for each of the 10 segments.

Sample selection was driven by the power to detect methylation differences between HPV16 infections (<CIN2) and precancers (CIN3), which required at least 12 specimens in each group to detect a minimum methylation difference of 0.25 with 80% power. For this discovery effort, we restricted to single HPV16 positive specimens to limit the influence of HPV type differences (approximately 12% of women who were managed by LEEP). From these, we drew a representative sample of 50 LEEP specimens taken from 33 women with available tissue blocks according to the sample size requirements described above. Areas of FFPE tissue sections on microscope slides that were representative of the underlying disease state, i.e. HPV16 infection, CIN1, CIN3, or cancer were microdissected for DNA extraction. Due to the limited DNA yields from individual microdissected tissue specimens, these samples were pooled according to diagnostic category and age group. Sixteen sample pools were constructed comprising DNA from 2 to 5 women each: 3 HPV16 infection pools (i.e., “normal”; n=10 samples), 4 CIN1 pools (n=13 samples), 5 CIN3 pools (n=13 samples), and 4 cancer pools (n=10 samples). An additional 4 cancer samples that were not pooled were also included.

Independent Sample Source for Candidate Biomarker Validation in Cytology Specimens

Validation of candidate gene markers was carried out in 206 LBC specimens obtained from a large, independent clinical laboratory. For this effort, we selected a stratified random sample of cytologic endpoints from women aged 20–70 years undergoing routine cervical cancer screening, irrespective of HPV infection. Cytologic diagnoses of these specimens included 70 (34%) negative for intraepithelial lesion or malignancy (NILM), 22 (11%) atypical squamous cells of undetermined significance (ASC-US) – including one sample with cannot exclude high-grade lesion (ASC-H), 47 (23%) low-grade squamous intraepithelial lesion (LSIL) and 67 (32%) high-grade intraepithelial lesion (HSIL). Sixty-six (32%) of the LBC specimens had concurrent histology results: 27 (41%) normal, 31 (47%) CIN1, 4 (6%) CIN2, 3 (5%) CIN3, and 1 (1%) squamous cell carcinoma (SCC). A total of 125 specimens were positive for 1 or more of the 13 carcinogenic HPV genotypes (6% of NILM, 79% of LSIL, and 94% of HSIL+).

Methylation Profiling Using a Bead-Based Array

Crude lysates were extracted from microdissected FFPE tissues using proteinase K digestion. Pooled DNA samples and DNA from the four individual cancer samples were treated with sodium bisulfite using the EZ DNA Methylation-Gold® Kit (Zymo Research, CA). Bisulfite treatment of DNA results in the deamination of unmethylated cytosine nucleotides to uracil while methylated cytosines remain unchanged²⁶. Following bisulfite conversion, methylation profiling was performed using the GoldenGate® Assay for Methylation and the standard Cancer Methylation Panel I (Illumina, Inc., CA), which includes 1,505 CpG sites proximal to the transcription start sites of 807 genes²⁷. Image processing and intensity data extraction were performed using BeadStudio Methylation Module software (Illumina, Inc., CA). The methylation status for each CpG site is presented as a β-value (ranging from 0 to 1), and is determined by computing fluorescent signals from the M (methylated) allele relative to that of the U (unmethylated) allele²⁷. Background signal generated by built-in negative controls is first used to normalize fluorescence data then normalized β-values are calculated by taking the ratio of the methylated signal intensity relative to the sum of both the methylated and unmethylated signals (M/(M+U)). β-value differences ≥0.17 were chosen as a minimum threshold for methylation differences²⁷. Methylation marker discovery analyses were performed using NCI Biometric Research Branch BRB-ArrayTools v4.2.1 and R v2.13.2 software.

Pyrosequencing of Candidate Genes

Pyrosequencing was performed on bisulfite-treated DNA from 206 LBC specimens. Pyrosequencing assays were developed for the top 12 candidate genes identified in the discovery microarray analysis (ADCYAP1, ASCL1, ATP10, CADM1, DCC, DBC1, HS3ST2, MOS, MYOD1, SOX1, SOX17, and TMEFF2 ) plus three additional candidate genes chosen from the literature (EPB41L3, MAL, miR124)^{5, 7, 8, 12, 14, 28}. We obtained sequences for all CpG allele-specific probes in the Cancer Methylation Panel I and designed flanking primers using PyroMark Assay Design Software 1.0 (Qiagen, CA) to those CpG sites for which a β-value was reported by Illumina. On average, about 9 CpG sites per gene were interrogated. Bisulfite conversion was performed on 500ng of extracted DNA using the EZ DNA Methylation-Gold® Kit (Zymo Research, CA) and eluted in 40 μl. Bisulfite-treated DNA (1–2 μl) was then amplified using HotStarTaq DNA Polymerase (Qiagen, CA) and the following PCR conditions: 95ºC for 15 min; 45 cycles of 95ºC for 30 sec, 50 – 62ºC for 30 sec, 72ºC for 30 s; 72ºC for 5 min; 4°C hold. Pyrosequencing reactions were performed on a PyroMark Q96 MD (Qiagen, MD) according to the manufacturer’s recommended protocol. Methylation status at each CpG site was measured by calculating the ratio of C (methylated cytosine) relative to T (unmethylated cytosine)²⁹. Individual CpG values were averaged to generate a mean percent methylation value for each gene. Raw data were analyzed using the Pyromark Q96 analysis software (Qiagen, MD).

Targeted, Next-Generation Bisulfite Sequencing of Candidate Genes

tNGBS was performed on 206 LBC specimens. A total of 15 pyrosequencing simplex assays were converted to tNGBS using multiplex PCR following library preparation with a fragment library kit (Kapa Biosystems, MA). The library-prepared multiplex PCR was templated using the Ion PGM™ Template OT2 200 Kit, and the sequenced using the Ion PGM™ Sequencing Hi-Q™ OT2 Kit with the Ion 314™ Chip Kit v2 on an Ion PGM™ System (Thermo Fisher Scientific, MA). FASTQ files from the Ion PGM™ System were filtered and aligned to the human genome assembly hg38 using Bismark Bisulfite Mapper v0.15.0 (Babraham Bioinformatics, UK) with the Bowtie 2 alignment algorithm. Methylation levels were calculated in Bismark by dividing the number of methylated reads by the number of total reads, considering all CpG sites covered by a minimum of 30 total reads.

Statistical Analysis

Methylation profiling data were visualized with unsupervised hierarchical clustering analysis using the Euclidean metric, and complete linkage clustering methods for the 100 most variable probes to characterize the distribution of methylation β-values in an unbiased fashion. To identify methylation markers for detection of cervical precancer and cancer in the discovery phase, we evaluated the distribution of DNA methylation levels for each CpG site across <CIN2, CIN3, and cancer, and conducted pairwise comparisons to identify genes that best distinguished between <CIN2 from CIN3. Methylation levels of individual CpG sites were compared using a t-test, and results were summarized and ranked by resulting p-values for CIN3 vs. CIN2. After evaluating all genes, we selected a group of 12 candidates (ADCYAP1, ASCL1, ATP10, CADM1, DCC, DBC1, HS3ST2, MOS, MYOD1, SOX1, SOX17, and TMEFF2) that met the following criteria: 1) increasing methylation with worse disease status (either CIN3 vs. <CIN2 or CIN3+ vs. <CIN2) 2) differences in methylation between CIN3 and <CIN2 >0.17, 3) mean methylation in <CIN2 <0.2, and 4) t-test p-value <0.1 for CIN3 vs. CIN2. We further prioritized genes with significantly different methylation levels between <CIN2 and CIN3 in two or more CpG sites and/or genes with any CpG site demonstrating highly significant differences in methylation (p<0.01) between <CIN2 and CIN3 or between <CIN2 and CIN3+. We also selected three additional candidate genes from the literature that were not included in the Illumina Cancer Methylation Panel (EPB41L3, MAL, miR124)^{5, 7, 8, 12, 14, 28} for validation with pyrosequencing and tNGBS analysis of LBC specimens. To reduce misclassification²⁵, combined endpoints were used based on cytology and histology results (when available) to create two disease categories: <CIN1 histology including those without a biopsy combined with <HSIL cytology (“<HSIL”, n=138, 67%) and CIN2+ histology combined with HSIL cytology (“HSIL+”, n=68, 33%). For the pyrosequencing and tNGBS results, Mann-Whitney rank sum tests were used to assess differences in median methylation levels for each gene comparing <HSIL to HSIL+. Spearman’s rho correlation coefficients were calculated for comparison of the methylation results obtained from pyrosequencing and tNGBS and to estimate the correlation of methylation levels between the 15 candidate genes. For each gene, we dichotomized methylation levels into high vs. medium/low tertiles based on the distribution in <HSIL samples, and calculated unadjusted and age-adjusted odds ratios (OR) and 95% confidence intervals (CI) to quantify the association between methylation at each gene and the odds of HSIL+. We used receiver operating characteristics (ROC) and areas under the curve (AUCs) with 95% CI’s to evaluate the ability of methylation at each gene to discriminate HSIL+ from <HSIL. For the top performing candidate gene markers (AUCs ≥0.75 for either method), we calculated the specificity at fixed sensitivity values of 80%, 85%, and 90%, respectively. Additionally, we analyzed different combinations of the top candidate gene markers with the highest AUCs, to evaluate whether a combined methylation marker panel could improve sensitivity and specificity beyond that achieved for an individual gene. For each gene included in the panel, we determined the cutoff value (% methylation) at a fixed sensitivity of 85% from the ROC curve. To maximize specificity, a sample was considered positive if the observed % methylation in all markers included in the panel was higher than the cutoff value established for each gene. We also verified the consistency of our results among the subset of HPV-positive LBC samples and performed a sensitivity analysis using pyrosequencing data from those samples with histologically confirmed <CIN2 (n=58) and CIN3 (n=8). All analyses were performed in Stata13 (Statacorp, College Station, TX).

Results

Discovery of Candidate Markers

DNA Methylation Profiling in FFPE Tissue Specimens

Unsupervised hierarchical clustering of DNA methylation profiling data from 1,505 CpG sites in 50 FFPE tissue samples from women with <CIN2, CIN3, and cancer revealed two distinct clusters with good separation of <CIN2 from CIN3 and cancers (Figure 1). Cluster 1 contained all of the <CIN2 samples while cluster 2 included the majority of cancer samples. Methylation levels were generally higher in cluster 2. Based on an assessment of % methylation differences in CIN3 vs. <CIN2, a total of 40 CpG sites within 30 genes met our criteria for candidate gene consideration (Supplementary Table 1). From these, we selected the following 12 candidate genes for the validation effort: ADCYAP1, ASCL1, ATP10, CADM1, DCC, DBC1, HS3ST2, MOS, MYOD1, SOX1, SOX17, and TMEFF2 (Table 1) plus an additional three identified from the literature (EPB41L3, MAL, miR124)^{5, 7, 8, 12, 14, 28}.

Figure 1.

Unsupervised hierarchical clustering of formalin‐fixed paraffin‐embedded cervical tissue samples based on the 100 most variable CpG sites using the Euclidean metric and the complete linkage clustering methods. CpG sites are in rows and cervical samples are in columns. Color coding of columns indicates cancer samples (green), CIN3 (pink) and HPV‐infected samples or CIN1 (blue). The heatmap shows the β‐values with darker blue colors indicating hypomethylation and yellow colors indicating hypermethylation.

Table 1.

DNA methylation differences in <CIN2, CIN3, and cancers in individual CpG sites within candidate gene markers identified in the Discovery Phase^{^} from 50 formalin-fixed paraffin-embedded tissue specimens

				Mean % Methylation			Difference (%)	t-test p-value
Gene	Illumina Probe ID (CpG Site)	Position*	Region	<CIN2	CIN3	Cancer	CIN3 vs. <CIN2	CIN3 vs. <CIN2	CIN3+ vs. <CIN2
SOX17	sox17_p287_r	55532761	Promoter	0.12	0.64	0.49	0.52	0.0016	0.0042
	sox17_p303_f	55532745	Promoter	0.10	0.50	0.43	0.40	0.0085	0.0032
CADM1	igsf4_p86_r	114880779	Promoter	0.04	0.33	0.27	0.29	0.0108	0.0232
	igsf4_p454_f	114880411	Promoter	0.0	0.03	0.06	0.03	0.0173	0.1531
TMEFF2	tpef_seq_44_s88_r	192767395	Exon	0.06	0.43	0.52	0.37	0.0195	0.0019
	tmeff2_p152_r	192768041	Promoter	0.05	0.40	0.42	0.35	0.0348	0.0057
	tmeff2_e94_r	192767795	Exon	0.02	0.29	0.26	0.27	0.0958	0.0958
	tpef_seq_44_s36_f	192767447	Exon	0.02	0.10	0.16	0.08	0.1311	0.0535
	tmeff2_p210_r	192768099	Promoter	0.06	0.22	0.14	0.16	0.2014	0.2140
DCC	dcc_p471_r	48120685	Promoter	0.13	0.52	0.59	0.39	0.0228	0.0036
	dcc_p177_f	48120979	Promoter	0.08	0.42	0.44	0.34	0.0321	0.0085
	dcc_e53_r	48121209	Exon	0.35	0.52	0.46	0.17	0.3088	0.2976
MOS	mos_e60_r	57189035	Exon	0.12	0.53	0.64	0.41	0.0245	0.0017
	mos_p27_r	57189122	Promoter	0.08	0.14	0.33	0.06	0.2854	0.0532
	mos_p746_f	57189841	Promoter#	0.63	0.80	0.48	0.17	0.1797	0.7990
ASCL1	ascl1_p747_f	101874847	Promoter	0.16	0.48	0.61	0.32	0.0376	0.0042
	ascl1_e24_f	101875618	Exon	0.01	0.27	0.40	0.26	0.1137	0.0112
ADCYAP1	adcyap1_p398_f	894989	Promoter	0.04	0.38	0.56	0.34	0.0479	0.0032
	adcyap1_p455_r	894932	Promoter	0.04	0.32	0.41	0.28	0.0803	0.0134
	adcyap1_e163_r	895550	Exon	0.02	0.11	0.10	0.09	0.2975	0.1996
SOX1	sox1_p294_f	111769620	Promoter	0.13	0.47	0.64	0.34	0.0483	0.0032
	sox1_p1018_r	111768896	Promoter	0.04	0.29	0.27	0.25	0.0620	0.0278
DBC1	dbc1_e204_f	121171318	Exon	0.09	0.37	0.53	0.28	0.0529	0.0077
	dbc1_p351_r	121171873	Promoter	0.05	0.34	0.51	0.29	0.0875	0.0087
HS3ST2	hs3st2_e145_r	22733506	Exon	0.12	0.46	0.60	0.34	0.0548	0.0040
	hs3st2_p171_f	22733190	Promoter	0.04	0.30	0.37	0.26	0.1268	0.0148
	hs3st2_p546_f	22732815	Promoter	0.09	0.31	0.29	0.22	0.1553	0.0488
MYOD1	myod1_e156_f	17697891	Exon	0.06	0.38	0.52	0.32	0.0564	0.0041
	myod1_p50_f	17697685	Promoter	0.01	0.14	0.06	0.13	0.0482	0.0765
ATP10A	atp10a_p147_f	23660110	Promoter	0.19	0.46	0.51	0.27	0.0688	0.0094
	atp10a_p524_r	23660487	Promoter	0.50	0.70	0.43	0.20	0.1573	0.8044

^{^}Candidate gene markers that met criteria are shown in bold.

^*Genome position according to NCBI 36.1/hg18 assembly

^#All CpG sites located within CpG islands except mos_p746_f

Abbreviations: CIN2/3, cervical intraepithelial neoplasia grade 2 and 3, respectively

Validation of Candidate Markers

Validation of the Top 15 Candidate Methylation Markers in LBC Specimens

Of the 206 LBC samples included in the validation effort, 39 did not pass multiplex PCR quality control assessment, and of those, 22 also had low signal/failures in at least one of the pyrosequencing assays. Thus, data are presented for 167 LBC specimens: 108 (78%) <HSIL, and 59 (87%) HSIL+. Of these, 105 (62.9%) were HPV-positive (n=49, 45% <HSIL and n = 56, 95% HSIL+, respectively)

In general, median methylation levels measured by pyrosequencing were similar to those measured by tNGBS, with Spearman correlation coefficients ranging from 0.34 – 0.91 (Table 2). All gene markers had significantly higher methylation levels in HSIL+ compared with <HSIL (p<0.05), irrespective of methylation detection assay. When we restricted the analysis to HPV-positive samples only (n=105), Spearman correlation coefficients comparing median methylation levels measured by pyrosequencing and tNGBS were similar to those observed for the whole study population, ranging from 0.45 to 0.91. Likewise, patterns were similar with respect to disease state, with all genes showing significantly higher methylation levels in HSIL+ compared with <HSIL (except for ATP10A [both methods], p<0.01, and CADM1 [pyrosequencing], p<0.1).

Table 2.

DNA methylation levels in 15 candidate gene markers in 167 LBC samples determined by pyrosequencing and tNGBS

	Pyrosequencing Median Methylation (%)		tNGBS Median Methylation (%)
Gene	<HSIL (n=108)	HSIL+ (n=59)	<HSIL (n=108)	HSIL+ (n=59)	Spearman’s Correlation Coefficient^
ADCYAP1	0.8	2.8	2.4	4.9	0.56
ASCL1	10.1	12.2	9.9	13.4	0.84
ATP10A	10.6	12.5	10.5	13.0	0.91
CADM1	0.6	1.0	0.5	0.8	0.34
DBC1	3.9	5.5	3.3	5.2	0.62
DCC	4.3	7.6	3.2	6.6	0.54
EBP41L3	1.0	5.9	0.6	2.7	0.61
HS3ST2	4.8	9.5	3.1	6.2	0.66
MAL	4.8	6.4	7.8	10.2	0.73
miR-124	9.0	11.8	8.8	12.5	0.80
MOS	8.2	11.9	7.3	10.2	0.68
MYOD1	11.6	13.9	7.5	9.7	0.65
SOX1	2.3	9.8	2.6	10.3	0.68
SOX17	4.1	5.8	4.5	7.6	0.66
TMEFF2	2.5	3.5	1.6	2.4	0.59

For all comparisons, all genes have significantly higher (Mann-Whitney p-value <0.05) methylation levels in HSIL+ compared with <HSIL.

^{^}Spearman’s rho calculated to assess the correlation between pyrosequencing and tNGBS

Abbreviations: LBC, liquid based cytology; tNGBS, targeted, Next-Generation Bisulfite Sequencing, HSIL, high grade squamous intraepithelial lesion

We calculated crude and age-adjusted ORs for each candidate gene, comparing methylation levels in HSIL+ to <HSIL. As shown in Table 3, unadjusted ORs ranged from 1.1 (95% CI: 0.6 – 2.2) for ATP10A to 16.2 (95% CI: 6.7 – 39.5) for SOX1 for pyrosequencing and from 1.1 (95% CI: 0.6 – 2.2) for ATP10A to 14.9 (95% CI: 6.4 – 35.1) for EPB41L3 for tNGBS. Results were similar for the age-adjusted analyses (data not shown). Among HPV-positive samples only, results were generally consistent, with unadjusted ORs ranging from 1.4 (95% CI: 0.6 – 3.1) for ATP10A to 12.0 (95% CI: 4.6 – 31.3) for SOX1 for pyrosequencing and from 1.1 (95% CI: 0.5 – 2.4) for ATP10A to 15.9 (95% CI: 5.9 – 43.1) for SOX1 for tNGBS (Supplementary Table 2).

Table 3.

Performance of 15 candidate gene markers in 167 LBC samples determined by pyrosequencing and tNGBS

	Pyrosequencing				tNGBS
Gene	OR	95% CI	AUC	95% CI	OR	95% CI	AUC	95% CI
ADCYAP1	7.6	3.6 – 16.0	0.75	0.67 – 0.84	5.3	2.7 – 10.7	0.79	0.71 – 0.86
ASCL1	2.6	1.4 – 5.1	0.69	0.60 – 0.78	3.3	1.7 – 6.5	0.72	0.64 – 0.80
ATP10A	1.1	0.6 – 2.2	0.59	0.49 – 0.68	1.1	0.6 – 2.2	0.61	0.52 – 0.70
CADM1	3.1	1.6 – 6.0	0.62	0.52 – 0.72	3.7	1.9 – 7.2	0.72	0.64 – 0.81
DBC1	3.6	1.8 – 7.3	0.71	0.62 – 0.79	3.3	1.7 – 6.5	0.75	0.67 – 0.83
DCC	10.1	4.6 – 22.2	0.82	0.74 – 0.89	8.8	4.1 – 19.2	0.82	0.75 – 0.89
EPB41L3	7.3	3.5 – 15.2	0.80	0.72 – 0.88	14.9	6.4 – 35.1	0.87	0.81 – 0.93
HS3ST2	6.2	3.0 – 12.7	0.78	0.70 – 0.86	8.5	4.0 – 18.1	0.82	0.75 – 0.90
MAL	3.3	1.7 – 6.5	0.73	0.65 – 0.81	3.3	1.7 – 6.4	0.69	0.60 – 0.77
miR-124	5.3	2.7 – 10.6	0.75	0.67 – 0.83	7.3	3.5 – 15.2	0.79	0.72 – 0.87
MOS	8.0	3.8 – 16.9	0.79	0.72 – 0.87	5.5	2.7 – 11.2	0.78	0.70 – 0.85
MYOD1	3.6	1.9 – 7.1	0.73	0.65 – 0.81	3.6	1.8 – 7.0	0.75	0.67 – 0.82
SOX1	16.2	6.7 – 39.5	0.86	0.79 – 0.92	11.3	5.0 – 25.5	0.85	0.78 – 0.92
SOX17	3.3	1.7 – 6.5	0.72	0.63 – 0.81	6.3	3.1 – 12.8	0.78	0.70 – 0.86
TMEFF2	3.4	1.7 – 6.5	0.71	0.63 – 0.80	3.2	1.7 – 6.2	0.71	0.62 – 0.79

ORs are calculated using univariate logistic regression; dichotomization of methylation levels is based on the distribution in <HSIL, comparing high vs. medium/low tertiles.

Abbreviations: tNGBS, targeted, Next-Generation Bisulfite Sequencing; OR, odds ratio, CI, Confidence Interval; AUC, area under the curve; HSIL, high grade squamous intraepithelial lesion

Clinical Performance of Candidate DNA Methylation Markers

We evaluated the ability of DNA methylation markers to discriminate HSIL+ from <HSIL cytology by calculating AUC values for each candidate gene marker. Using pyrosequencing, eight gene markers had an individual AUC value of ≥0.75 (ADCYAP1, DCC, EPB41L3, HS3ST2, miR-124, MOS, SOX1, TMEFF2) and using tNGBS, nine markers had an AUC value of ≥0.75 (ADCYAP1, DBC1, DCC, EPB41L3, HS3ST2, miR-124, MOS, SOX1, SOX17) (Table 3). Overall, SOX1, DCC, and EPB41L3 showed the best discrimination with AUC values of ≥0.80, irrespective of methylation detection assay. Based on tNGBS results, HS3ST2 also showed promising discrimination of HSIL+ from <HSIL with an AUC of 0.82. Restricting to only HPV-positive samples, SOX1 and DCC showed the best discrimination of HSIL+ from <HSIL, with AUC values of 0.85 and 0.79 for pyrosequencing and 0.84 and 0.80 for tNGBS, respectively. Based on tNGBS results, EPB41L3 and HS3ST2 also showed promising discrimination of HSIL+ with AUCs of 0.85 and 0.82, respectively (Supplementary Table 2). Among the subset of samples with available histology data (n=66 for pyrosequencing), DCC showed the best discrimination with an AUC of 0.85. All top candidates identified in our main analysis (SOX1, EPB41L3, and HS3ST2) were among the top performing markers in this smaller subset, with AUC’s >0.70 (Supplementary Table 3).

Performance of Candidate Methylation Markers at Fixed Sensitivities

Based on the assumption that a triage test would require high sensitivity for the detection of CIN3+, we examined the performance of the top candidate gene markers at fixed sensitivity values of 80%, 85%, and 90%, respectively (Table 4). At a fixed sensitivity of 80%, the specificity for the top candidate markers was 77% for SOX1, 62–71% for DCC, and 60–84% for EPB41L3, depending on the methylation detection assay. Values of specificity were slightly reduced at a fixed sensitivity of 85%, ranging from 67–72%, 50–57%, and 40–76% for SOX1, DCC, and EPB41L3, respectively and at fixed sensitivity of 90%, ranging from 52–59%, 32–38%, and 24–57% for SOX1, DCC, and EPB41L3, respectively. Restricting to HPV-positives yielded similar results. For example, at a fixed sensitivity of 80% the specificity of top candidates ranged from 73–80%, 57–67%, and 63–82% for SOX1, DCC, and EPB41L3, respectively (data not shown).

Table 4.

Specificity at fixed sensitivities of top candidate gene markers in 167 LBC samples determined by pyrosequencing and tNGBS

Gene	Sensitivity = 80%		Sensitivity = 85%		Sensitivity = 90%
	Pyrosequencing	tNGBS	Pyrosequencing	tNGBS	Pyrosequencing	tNGBS
ADCYAP1	60%	58%	35%	55%	<25%	50%
DBC1	42%	41%	32%	39%	21%	33%
DCC	71%	62%	57%	50%	32%	38%
EPB41L3	60%	84%	40%	76%	24%	57%
HS3ST2	57%	63%	36%	52%	31%	29%
miR-124	47%	60%	39%	59%	28%	36%
MOS	67%	57%	38%	55%	25%	32%
SOX1	77%	77%	72%	67%	59%	52%
SOX17	38%	58%	24%	39%	21%	23%
TMEFF2	45%	43%	33%	41%	22%	30%

Abbreviations: LBC, liquid based cytology; tNGBS, targeted, Next-Generation Bisulfite Sequencing

Performance of a Combined Panel of Candidate Methylation Markers

To assess whether a panel of methylation markers could improve sensitivity and specificity beyond that achieved for an individual gene, we conducted an exploratory analysis of different combinations of top candidate markers, also taking into account the observed correlation coefficients for methylation levels between top candidates (Supplementary Table 4). A combined panel of DCC, EPB41L3 and SOX1, considered positive if methylation in all three genes was higher than the established cutoff for each gene, achieved a sensitivity and specificity of 70% and 82% for pyrosequencing and 70% and 91% for tNGBS, respectively. Using tNGBS, pairwise combinations of DCC, EPB41L3 and SOX1 resulted in sensitivities and specificities of 76% and 87% (EPB41L3 and SOX1), 76% and 86% (EPB41L3 and DCC) and 75% and 84% (SOX1 and DCC), respectively. Pairwise Spearman correlation coefficients for these markers ranged from 0.53 for SOX1 and DCC to 0.64 for SOX1 and EPB41L3 (Supplementary Table 4). Restricting to HPV-positives, the three gene panel achieved a sensitivity and specificity of 73% and 77% for pyrosequencing and 68% and 90% for tNGBS, respectively (data not shown).

Discussion

We carried out a DNA methylation-profiling study to identify host gene DNA methylation markers associated with cervical precancer and cancer in tissue and validated these findings in an independent population of cytology specimens, using high-resolution bisulfite sequencing methods. Based on our discovery effort, we identified 12 candidate gene markers (ADCYAP1, ASCL1, ATP10, CADM1, DCC, DBC1, HS3ST2, MOS, MYOD1, SOX1, SOX17, and TMEFF2) that showed significantly elevated methylation in DNA from women with CIN3+ versus women with <CIN2. Subsequent validation of these top candidates as well as EPB41L3, MAL, and miR-124 in LBC samples using pyrosequencing and tNGBS revealed a total of seven markers with high diagnostic accuracy for HSIL+, irrespective of methylation detection assay (ADCYAP1, DCC, EPB41L3, HS3ST2, miR-124, MOS, SOX1), with SOX1, DCC, EPBL143, and HS3ST2 showing the greatest promise based on AUCs ≥0.80 for one or both assays. At a fixed sensitivity of 85%, the performance of our top candidate markers was similar to other existing molecular triage strategies, such as immunostaining with p16INK4a^30–32 and HPV DNA methylation in multiple carcinogenic types³³.

In line with other studies, we identified well-established candidate markers, including SOX1, miR-124, and EPB41L3^{4, 10–12, 14, 28, 34}, which supports the validity of our findings. Several of the more novel markers that we identified have been previously associated with other cancers and/or are involved in important pathways related to carcinogenesis (HS3ST2³⁵, ADCYAP1³⁶, DCC^37–39, MOS⁴⁰, SOX17⁴¹, and TMEFF2⁴²). To our knowledge, our study is the first to identify DCC as a top candidate methylation marker for cervical cancer screening. DCC is a tumor suppressor gene involved in cell cycle control and apoptosis, and has been implicated in several cancers, including colorectal, esophageal, and ovarian cancer^{37–39, 43, 44}. Interestingly, DCC methylation has been associated with head and neck squamous cell carcinoma⁴⁵ and in one study, was found to be differentially methylated in head and neck cancer cell lines by HPV status⁴⁶. In a recent study of women undergoing routine cervical cancer screening in Appalachian Ohio, pyrosequencing of candidate genes revealed an association of DCC methylation with increased odds of abnormal cytology⁴⁷. Further studies are needed to determine whether methylation of DCC can serve as a triage marker for cervical cancer screening.

We also noted some important differences in our results compared with findings from other studies. For example, multiple studies using qMSP-based assays have shown that CADM1, in combination with MAL, holds promise as a potential triage tool for HPV-positive women^{5–9, 16, 48}; however, we were unable to validate these results in our study. In contrast, our findings are similar to a recent study conducted by Vasiljevic et al., in which the methylation levels of 26 candidate genes were evaluated using pyrosequencing¹⁰. In their study, median methylation values of CADM1 (1.6 % in <CIN2 and 2.6% in CIN2/3) and MAL (2.4 % in <CIN2 and 3.2% in CIN2/3) were relatively similar to those observed in our study using pyrosequencing. Further, among their top candidate genes assessed, CADM1 and MAL had the lowest AUC values (0.58 and 0.62, respectively)¹⁰. These discrepancies may be explained by important methodological differences between methylation detection assays as discussed by Wentzensen et al³. For example, it is possible that while the small absolute differences in CADM1 methylation we observed using pyrosequencing or tNGBS were not large enough to discriminate HSIL+ from <HSIL, these differences may be amplified in studies utilizing qMSP¹⁰. Additionally, some of the candidate regions and/or genes interrogated in our study may not correspond to the same CpG sites previously identified. These discrepancies emphasize the importance of validating findings with different technologies in different study populations, and in different laboratory settings⁴⁹. Indeed, while qMSP is highly sensitive for detection of the low levels of DNA methylation that are present in cytology samples, these assays are typically designed only to interrogate a small number of CpG sites, and the range of detection sensitivities and options for different positivity thresholds make it more challenging to standardize across different studies and/or laboratory settings⁴⁹. In contrast, methods such as pyrosequencing and tNGBS, provide increased genomic coverage and allow for absolute quantitation of DNA methylation; however, the sensitivity limit of pyrosequencing (5%) may restrict its utility for detection of low-level DNA methylation in heterogeneous cytology samples⁵⁰.

Strengths of our study include the objective, microarray-based approach for marker discovery in FFPE tissue specimens, which enabled us to discern biologically relevant genes that may play an important role in the epigenetic events associated with cervical carcinogenesis. Further, we have demonstrated the ability of pyrosequencing and tNGBS to identify successfully DNA methylation markers with low background levels in cytology specimens, which is crucial for clinical assay development⁴⁹. While we consider our objective approach a strength of our study, it is important to note that the Illumina Cancer Methylation Panel I used for marker discovery in our study is purposely enriched for genes that are strongly implicated in cancer; thus we were unable to verify certain candidate markers (e.g., FAM194A^{13, 15}) and acknowledge that additional candidates may have been identified if we had used an alternative DNA methylation profiling platform with broader coverage²⁷. In addition, since women included in our discovery phase were referred to colposcopy as a result of an abnormal Pap test or for treatment of already-diagnosed cancer, our results may not reflect differences that would be seen in a general screening population. With respect to the validation phase, we were unable to assess the performance of methylation markers in relation to histological outcome for all cytology samples, thus it is possible that there may have been disease misclassification using endpoints based on cytology alone. Finally, we had limited ability to account for potentially important confounders in our analyses, and we were not statistically powered to adjust for multiple comparisons in the discovery phase. Nevertheless, we were able to validate the top candidate markers in cytology specimens, and our results were consistent with those previously reported in other study populations using different assays for several candidate genes, lending support for the generalizability of our findings, and the robustness of our results to residual confounding factors and spurious associations.

In conclusion, our tissue-based biomarker discovery effort identified host DNA methylation markers associated with cervical cancer and precancer. We validated these findings in an independent population of LBC samples using two different high-resolution DNA-methylation detection methods. In addition to verifying candidate DNA methylation markers, our strategy identified several novel candidate markers for detection of cervical precancer in cytology specimens that warrant further validation in prospective studies.

Novelty and Impact:

We conducted a rigorous, tissue-based DNA methylation profiling study for biomarker discovery with validation of promising candidates in an independent set of cytology samples using pyrosequencing and next-generation bisulfite sequencing. Of the 15 candidates, 10 achieved an area under the curve of ≥ 0.75 for discrimination of high-grade from normal cytology. In addition to verifying candidate markers from the literature (e.g., SOX1 and EPB41L3), we identified several novel markers that warrant validation in prospective studies.

Abbreviations:

AUC

area under the curve

CIN

cervical intraepithelial neoplasia

FFPE

formalin-fixed paraffin-embedded

HPV

Human papillomavirus

HSIL

high grade squamous intraepithelial lesion

LBC

liquid based cytology

tNGBS

targeted next-generation bisulfite sequencing